AI Video Skill — Gerador com Aprendizado Contínuo
0xadvait/ai-video-skill
Geração de vídeo IA que aprende e se auto-melhora a cada clip
Ver no GitHubSkill Python para Claude Code que gera vídeos end-to-end através de 6 modelos (Seedance 2.0, Kling, Wan, Veo, OmniHuman) com loop automático de controle de qualidade. A ferramenta não apenas cria clips, mas avalia cada geração, aprende com os erros e melhora seus prompts iterativamente.
Instalação
npx convez-ai 0xadvait/ai-video-skill1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Produção de vídeos comerciais e marketing com qualidade controlada automaticamente
- Criação iterativa de conteúdo visual que se refina a cada geração
- Automação de pipeline completo: frame inicial → prompt → síntese de voz → montagem → upscale
Veredito

Ferramenta robusta que vai além da geração simples: implementa QC automático e aprendizado iterativo. Exige familiaridade com APIs de IA de vídeo e Python avançado, mas entrega pipeline production-ready com capacidade de melhoria contínua.
README
A tradução para o português deste README ainda está na fila.
ai-video 🎬
The AI video skill that learns.
End-to-end video generation for Claude Code — across six models, from prompt to finished cut. It quality-controls every clip it makes, and gets better at prompting with each one.
Most AI-video tooling stops at "here's your clip." ai-video doesn't. Every generation ends with a quality-control pass — Claude watches the rendered clip, scores it against the prompt, and writes down what it learned. That lesson is loaded back in before the next prompt is written. The skill is not static; it compounds.
It also covers the whole job, not just the model call: generate a first frame, write the prompt, run it on the right model, synthesize voiceover, stitch the clips, upscale for delivery.
Showcase
Real clips, real prompts — each generated with the structure this skill teaches. Open a prompt to see the 5-part craft.
[00:00-00:05] Extreme wide IMAX 70mm desert shot in the Denis Villeneuve and
Greig Fraser register: a tiny armored convoy races across hard salt flats
while a mile-high sandstorm advances behind it, desaturated ochre palette,
enormous negative space.
[00:05-00:10] Cut inside the lead rover, medium close-up of the driver
gripping the wheel at ten and two, dashboard amber reflected in the visor.
Camera shakes with engine vibration. He says, "Do not slow down." Sand
hammers the windshield.
[00:10-00:15] Exterior low tracking shot as the rover crests a dune in real
time, not slow motion. Dust swallows the frame, distant bass rumble, no gore,
no text, hard cut to black.
[00:00-00:05] Interior cockpit medium close-up of a veteran race driver at
night, rain lashing the windshield, dashboard LEDs reflected in the visor. He
tightens both gloved hands on the wheel at ten and two, breathing steady.
[00:05-00:10] Cut to rival cockpit, younger driver in the next car, jaw
tense, eyes forward, engine vibration shaking the frame. He whispers, "Hold
the line."
[00:10-00:15] Low exterior tracking shot as the green light hits and both
cars accelerate in real time on wet asphalt. Massive water spray hits the
lens, stadium lights stretch into motion blur, engines roar, no limb
distortion, no text.
Extreme macro close-up of a steak hitting a black cast-iron skillet,
background fully blurred. Fat renders, bubbles along the crust, butter foams
around rosemary and garlic, steam blooms into warm tungsten light. Slow
dolly-in, probe-lens feel, razor-thin depth of field. Audio: aggressive
sizzle, fat crackle, one tong clink at end, no music, no voice, no text.
Loop-ready studio macro shot of a perfect liquid mercury sphere on a black
mirror surface. Locked camera, centered composition, cool white strip lights
reflected across the metal. Over 6 seconds the sphere slowly deforms into a
rounded cube under an invisible force, then returns exactly to the original
sphere shape by the final frame. First and last frames match for seamless
playback. No text, no logo, no extra objects, no camera movement.
All four prompts — and 26 more — ship in examples/prompts.json.
🎞️ See the full results in motion — the complete clip library, with video previews and filterable prompts, lives at ai-video-guide.vercel.app/library.html.
Features
Install
git clone https://github.com/0xadvait/ai-video-skill.git ~/.claude/skills/ai-video
pip install -r ~/.claude/skills/ai-video/scripts/requirements.txt
ffmpeg + ffprobe must be on PATH (frame QC, stitching, upscaling). The
skill activates automatically — just ask Claude Code to make a video.
Heads up: API keys are read from the environment. You only need a key for the models you actually call (see Environment).
The pipeline
A single clip needs only stages 2–4. A finished video runs the whole chain.
| # | Stage | Script | What it does |
|---|---|---|---|
| 1 | Still | imagegen.py | First-frame / reference images — Flux Schnell, Flux Pro, Imagen 4 |
| 2 | Prompt | SKILL.md + reference/ | 5-part structure (Subject→Action→Camera→Style→Constraints) + time-coded blocks |
| 3 | Generate | validate.py → generate.py | Validate the request, run on the auto-routed model |
| 4 | QC + learn | review.py → LESSONS.md | Watch the clip, score it, record a prompting lesson |
| 5 | Voiceover | tts.py | ElevenLabs text-to-speech narration |
| 6 | Assemble | assemble.py | Stitch clips — cuts/crossfades + audio bed — from a JSON edit list |
| 7 | Finish | upscale.py | Upscale resolution + interpolate frame rate |
Models
One canonical schema (Seedance 2.0's). generate.py translates it to whichever
model is chosen — "model": "auto" (default) routes by intent.
| Model | Backend | Best for |
|---|---|---|
seedance-2.0 | Replicate | Generalist — quad-modal refs, native audio, lip-sync, motion transfer |
kling-3.0-omni | fal | t2v / i2v with a different moderation gateway (the --fallback) |
wan-2.7-i2v | Replicate | Open-weights image-to-video |
veo-3.1 | Google Gemini | Highest-realism cinematic text-to-video |
omnihuman-1.5 | fal | Audio-driven talking avatar, gesture-rich |
veed-fabric-1.0 | fal | Mouth-only lip-sync — included, but see LESSONS.md |
The learning loop
This is the part that makes it more than a wrapper:
prompt ──▶ generate ──▶ review.py ──▶ Claude watches the clip
▲ │
│ ▼
LESSONS.md ◀────────────── distill a generalizable lesson
review.py extracts a frame contact sheet; Claude scores the clip on prompt
fidelity, consistency, motion, and audio; a one-line cause→effect lesson is
appended to LESSONS.md. Step 0 of every workflow reads LESSONS.md back in.
No regeneration is needed for the skill to improve — clips are only re-run
when you ask.
Documentation
| File | What's inside |
|---|---|
SKILL.md | The workflow, model routing, QC loop, hard rules |
LESSONS.md | Seed lessons + the skill's growing memory |
reference/schema.md | Canonical Seedance 2.0 request schema |
reference/prompt-logic.md | 10 prompt categories, theory + perfect examples |
reference/style-library.md | Cross-model camera / lens / lighting / director vocabulary |
reference/real-person.md | Real-person workflow + production stacks |
reference/failure-modes.md | 15 failure modes and their fixes |
examples/prompts.json | 30 verified, paste-ready prompts |
Environment
| Variable | Used by |
|---|---|
REPLICATE_API_TOKEN | seedance-2.0, wan-2.7-i2v, Flux stills, ML upscale |
FAL_KEY | kling-3.0-omni, veed-fabric-1.0, omnihuman-1.5 |
GEMINI_API_KEY | veo-3.1, Imagen 4 stills |
ELEVENLABS_API_KEY | tts.py voiceover |
Personal profile
The "a video of me" shortcut uses a personal profile. Copy the template and fill in your own assets:
cp profile/profile.example.json profile/profile.json
profile.json is git-ignored — your avatar paths and voice id never leave
your machine.
A note on real-person video
Real-person generation is policy-sensitive. Only build around public
figures in legitimate commentary or satire, or consenting, tagged
collaborators — never undisclosed impersonation of private individuals.
Respect voice-provider restrictions when sourcing reference audio. See
reference/real-person.md.
Contributing
Issues and PRs welcome — new verified prompts for examples/prompts.json,
model adapters in scripts/generate.py, and reference-doc corrections are all
fair game. Keep prompt examples real (rendered, not invented).
License
MIT © 2026 Advait Jayant.
Built from the ai-video-guide research corpus.
🎬 Generated with Claude Code